AZ-965 ships the NetVLAD .pt checkpoint that clears the AZ-839
empty-c10_provisioning.backbones SKIP gate. Pipeline-integration
scaffold — encoder is real, NetVLAD tail is honestly labelled as
untrained.
Composition:
* Encoder (26 keys, encoder.0..encoder.28): torchvision
vgg16(weights=IMAGENET1K_V1) features [:-2], BSD-3-Clause.
Real ImageNet-pretrained VGG16 conv stack.
* NetVLAD pool + PCA tail (5 keys: pool.conv.{weight,bias},
pool.centroids, pca.{weight,bias}): random-init via
torch.manual_seed(0). NOT trained for visual place recognition.
Total: 149,002,112 params (568.4 MiB fp32, sha256=745c6f29...).
Round-trip verified locally: torch.load(weights_only=True) +
load_state_dict(strict=True) succeed; forward(1,3,480,480) emits
{'vlad_descriptor': (1, 4096) fp32} — matches NetVladStrategy
contract per net_vlad.py:247-251.
Two material discoveries documented in the AZ-965 spec:
1. The NetVLAD-VGG16 architecture already lives in repo at
src/gps_denied_onboard/components/c2_vpr/_net_vlad_architecture.py
— we instantiate it and save a state_dict, NOT externally source.
2. The PyTorch FP16 runtime expects a .pt state_dict (NOT .onnx).
BackboneConfig.onnx_path is a misnomer for NetVLAD: per AZ-321
design + c2_vpr description.md §1, NetVLAD runs on PyTorch FP16
(NOT TRT). compile_engine is a no-op sha256+path wrap;
deserialize_engine does torch.load(weights_only=True) +
load_state_dict(strict=True).
User skipped Option A/B/C/D/E question — judgment call = Option B
(IMAGENET1K_V1 + random tail) per "use judgment, don't block":
* Option A (Nanne translation) was 5-8 SP, above the 5 SP budget.
* Option B is 3 SP, fits the budget, honestly labelled.
* Option C (pure random) was borderline-dishonest per Real Results.
Files:
* scripts/mk_netvlad_checkpoint.py — deterministic generator.
* models/netvlad/netvlad.pt — 568 MiB, via git-lfs (.gitattributes
extended for models/**/*.pt, *.onnx, *.engine).
* configs/operator_replay.yaml — c2_vpr + c10_provisioning blocks
populated; the field literally named onnx_path actually points
at the .pt for NetVLAD per the runtime semantics noted above.
* docker-compose.test.jetson.yml — ./models:/opt/models:ro bind
mount added to e2e-runner.
* _docs/03_ip_attribution/netvlad.md — provenance, licence, how-to-
reproduce, honest scope statement ("NOT a real-retrieval
checkpoint; ESKF divergence under garbage retrievals is the
expected next gate").
* _docs/02_tasks/todo/AZ-965_netvlad_onnx_backbone_provisioning.md
— rewritten to reflect the .pt-not-.onnx + Option B discoveries.
Tier-2 verification follows in a separate commit after the harness
run confirms the empty-backbones SKIP gate clears.
Out of scope (filed as follow-ups):
* Real-retrieval NetVLAD weights (Nanne Pittsburgh-30k translation
or internal team checkpoint) — separate ticket.
* AZ-840 orchestrator PASSing end-to-end (depends on retrieval
quality + ESKF stability).
* AZ-963 60s smoke ESKF divergence (independent chain).
Co-authored-by: Cursor <cursoragent@cursor.com>
Oleksandr Bezdieniezhnykh